Generating new ideas with non-deterministic code or traversing nodes in a semantic network or using an LLM is relatively easy. The hard part is to get the artificial intelligence to understand if the idea is of value to humans. How do you plan to get the intelligence to focus on ideas that are of value to humans?
The background for this line of question is that in the 50’s and 60’s computers were programmed to explore math and solve proofs. While they were able to make progress they essentially want off with no real direction and soon were creating results that were of no real use and/or not able to return a result in a reasonable amount of time (think a day of processing)
“A Computing Procedure for Quantification Theory” by Martin Davis and Hilary Putnam. (pdf)
“The Early History of Automated Deduction” by Martin Davis (pdf)
For a simple example let’s get the system to just put two words together to form an idea and then tell us how the idea is of value to a human.
If the two words are “barrel” and “wheel” that gives “barrel wheel” or “wheel barrow” but the first thought without knowing what as human we currently understand as a wheelbarrow might be a barrel with a wheel attached at the top, never mind the thought of adding handles.
If the system reaches the state with the two words, can it progress on that idea by putting the wheel under the barrel, adding some handles and making it ergonomically useful to a human?